Design and Implementation Experiments of Scalable Socket Buffer Tuning

1. Design and Implementation Experiments ofScalable Socket Buffer Tuning Tatsuhiko Terai Department of Informatics and Mathematical Science, Graduate School of Engineering Science, Osaka University 4th Asia-Pacific Symposium on Information and Telecommunication Technologies

2. Background • Explosive increase of network traffic due to rapid increase of Internet users • Many improvements on network to accommodate increasing traffic • Link Bandwidth • TCP congestion control algorithm • Few discussions on endhost improvement • Ex. Busy WWW Server receives hundreds of request for document transfer per second. The bottleneck of the data transfer processing is shifted from network to endhosts 4th Asia-Pacific Symposium on Information and Telecommunication Technologies

3. Objective • Effective allocation of endhosts resources becomes more important • Propose a novel architecture, SSBT (Scalable Socket Buffer Tuning) • High-performance and fair service for many TCP connections at the sender host • E-ATBT (Equation-based Automatic TCP Buffer Tuning) • SMR (Simple Memory-copy Reduction) 4th Asia-Pacific Symposium on Information and Telecommunication Technologies

4. Send Socket Buffer Assignment • Busy Internet servers (WWW, Proxy,…) handle many TCP connections which have different characteristics • RTT, packet loss rate, bandwidth, … • In original method, sender host assigns fixed size of buffer to each TCP connection • Fixed size assignment may cause unfair and ineffective usage of send socket buffer • Different connections require different size of send buffer according to network conditions Fair and effective buffer assignment considering network condition is needed for improving endhost performance 4th Asia-Pacific Symposium on Information and Telecommunication Technologies

5. E-ATBT(Equation-based Automatic TCP Buffer Tuning) • Provide fair and effective send socket buffer assignment • Estimate an ‘expected’ throughput of each TCP connection by monitoring three network • p (packet loss rate), rtt (Round Trip Time), rto (Retransmission Time Out) • Determine assigned buffer size from the estimated throughput • Max-Min fairness policy for re-assigning the excess buffer • Re-assigned to the connections need more buffer the required buffer size of those connection 4th Asia-Pacific Symposium on Information and Telecommunication Technologies

6. E-ATBT Method for assigning the send socket buffer In E-ATBT, assigned buffer size is determined from the estimated throughput based on the mathematical analysis method Web Server Host Internet 4th Asia-Pacific Symposium on Information and Telecommunication Technologies

7. E-ATBT Method for assigning the send socket buffer In E-ATBT, assigned buffer size is determined from the estimated throughput based on the mathematical analysis method Web Server Host Assign the large socket buffer for large bandwidth Internet Assign the required buffer size only Send Socket Buffer TCP Connections 4th Asia-Pacific Symposium on Information and Telecommunication Technologies

8. Application Buffer Socket Buffer Data File SMR (Simple Memory-copy Reduction) • Original mechanism needs two memory copy operations. Memory copy is large overhead on endhost processing Main Memory File System Memory Copy Memory Copy 4th Asia-Pacific Symposium on Information and Telecommunication Technologies

9. Socket Buffer Data File SMR (Simple Memory-copy Reduction) File System Main Memory Memory Copy SMR scheme reduces the data transfer overhead at the sender host by reducing the redundant memory copy operation 4th Asia-Pacific Symposium on Information and Telecommunication Technologies

10. Implementation Experiments • Fair buffer assignment among different connections • Time-dependent behavior of the assigned buffer size • Web server performance evaluation • Average performance gain of the SSBT scheme • Considering realistic web access traffic [1] • Document size distribution • Idle time distribution of requests • Embedded documents distribution [1] P.Barford and M.Crovella, “Generating Representative Web Workloads for Network and Server Performance Evaluation”, in Proceedings of ACM SIGMETRICS ’98, 1998 4th Asia-Pacific Symposium on Information and Telecommunication Technologies

11. Network topology Document transfer request generated by httperf Router 1 1000 [Mbps] SSBT-enabled Web server 1000 [Mbps] Client Host 1 100 [Mbps] Server Host 100 [Mbps] Router 2 Client Host 2 Each client generate the requests for document transfer to Web server and measure the data transfer time 4th Asia-Pacific Symposium on Information and Telecommunication Technologies

12. 600000 Connection 1 Connection 2 Connection 3 500000 Connection 4 400000 Assigned Send Socket Buffer [Byte] 300000 200000 100000 0 0 5 10 15 20 25 30 35 40 Transfer Time [s] Buffer assignment result E-ATBT can provide the stable and fair buffer assignment 4th Asia-Pacific Symposium on Information and Telecommunication Technologies

13. Web server performance 600Clients, HTTP/1.1 1000 500 300 200 Average Performance Gain [%] 100 Performance improvement by SSBT 50 30 20 10 10 100 1000 10000 Document Size [KBytes] 4th Asia-Pacific Symposium on Information and Telecommunication Technologies

14. Conclusion Remarks and Future Works • Proposed SSBT for utilizing the send socket buffer effectively and fairly • Confirmed the effectiveness of the SSBT algorithm through implementation experiments and shown SSBT can improve the overall performance of a server • New resource management scheme for Internet busy server (Ex. HTTP Proxy server) • Enhanced E-ATBT for proxy servers • Manage the persistent TCP connections 4th Asia-Pacific Symposium on Information and Telecommunication Technologies

15. 4th Asia-Pacific Symposium on Information and Telecommunication Technologies